Request: TensorFlow.Lite Flex delegate

[nexxuno]

Are you planning on supporting the Flex deleagte for tensor flow lite? It's needed by the interpreter to correctly load models from microsoft customvision's exported models for example.

[moljac]

@nexxuno

I'm not familiar with Flex delegate. Maybe some links. Is that separate lib?

Minimal sample, please?

It would help us/me to bind it faster and test it.

[nexxuno]

@moljac Is this of any help?

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/java/src/test/java/org/tensorflow/lite/InterpreterFlexTest.java

  public void testFlexModel() throws Exception {
    FlexDelegate delegate = new FlexDelegate();
    Interpreter.Options options = new Interpreter.Options().addDelegate(delegate);
    try (Interpreter interpreter = new Interpreter(FLEX_MODEL_BUFFER, options)) {
      testCommon(interpreter);
    } finally {
      delegate.close();
    }
  }

[moljac]

@nexxuno I have decompiled version 2.2.0 and there is no FlexDelegate

Google search: tensorflow flexdelegate

Results in this:

https://www.tensorflow.org/lite/guide/ops_select

Caution: This feature is experimental.

We do bindings for previews in some exceptional cases and I'm not sure this will be such. Sorry.

Please close this one and reopen new issue, when FlexDelegate comes out as stable.

[nexxuno]

@moljac thanks.

The problem is that the tf lite model exported with microsoft's custom vision service needs this kind of experimental feature to work. The custom vision team said they would try and remove this need from their model (see #54529 linked above) but there's no guarantee. If they'll keep using this functionality then I think it should be considered for integration even if it's in preview. Could you try and understand with them how to get out of the deadlock?